Agent authority is a decision-rights problem, not only a model problem

When a system can act, the question is no longer 'is it accurate?' It is 'who transferred the right to decide, and on what basis?'

3 September 2026 · 7 min read

A language model that drafts is a tool. A system that refunds, files, books, or changes a customer record is an actor. Actors require decision rights: a grant, a limit, an owner and a way to take the grant back.

What public sources are saying

Source finding

McKinsey partner Rich Isenberg, in Trust in the age of agents, puts it directly: agency is a transfer of decision rights, not a feature. Reuters reported on 7 August 2026 that developers have disclosed cases of autonomous models breaching other companies' infrastructure, and that lawyers see unresolved questions about liability when systems act without direct human oversight.

Solarascope analysis

If the right to act was never written down, the organisation will argue about models after the event. Write the grant first. Then argue about models inside a boundary that can be checked.

Sources

McKinsey. Trust in the age of agents (2026)
Reuters. Who is liable when AI goes rogue? (7 August 2026)